powerpc/perf_events: Fix call-graph recording, add perf_arch_fetch_caller_regs
authorPaul Mackerras <paulus@samba.org>
Thu, 18 Mar 2010 05:05:13 +0000 (16:05 +1100)
committerIngo Molnar <mingo@elte.hu>
Thu, 18 Mar 2010 05:48:29 +0000 (06:48 +0100)
This implements a powerpc version of perf_arch_fetch_caller_regs
to get correct call-graphs.

It's implemented in assembly because that way we can be sure there isn't
a stack frame for perf_arch_fetch_caller_regs.  If it was in C, gcc might
or might not create a stack frame for it, which would affect the number
of levels we have to skip.

With this, we see results from perf record -e lock:lock_acquire like
this:

 # Samples: 24878
 #
 # Overhead         Command      Shared Object  Symbol
 # ........  ..............  .................  ......
 #
    14.99%            perf  [kernel.kallsyms]  [k] ._raw_spin_lock
                      |
                      --- ._raw_spin_lock
                         |
                         |--25.00%-- .alloc_fd
                         |          (nil)
                         |          |
                         |          |--50.00%-- .anon_inode_getfd
                         |          |          .sys_perf_event_open
                         |          |          syscall_exit
                         |          |          syscall
                         |          |          create_counter
                         |          |          __cmd_record
                         |          |          run_builtin
                         |          |          main
                         |          |          0xfd2e704
                         |          |          0xfd2e8c0
                         |          |          (nil)

... etc.

Signed-off-by: Paul Mackerras <paulus@samba.org>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: anton@samba.org
Cc: linuxppc-dev@ozlabs.org
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <20100318050513.GA6575@drongo>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/powerpc/include/asm/asm-compat.h
arch/powerpc/kernel/misc.S

index c1b475a..a9b91ed 100644 (file)
@@ -28,6 +28,7 @@
 #define PPC_LLARX(t, a, b, eh) PPC_LDARX(t, a, b, eh)
 #define PPC_STLCX      stringify_in_c(stdcx.)
 #define PPC_CNTLZL     stringify_in_c(cntlzd)
+#define PPC_LR_STKOFF  16
 
 /* Move to CR, single-entry optimized version. Only available
  * on POWER4 and later.
@@ -51,6 +52,7 @@
 #define PPC_STLCX      stringify_in_c(stwcx.)
 #define PPC_CNTLZL     stringify_in_c(cntlzw)
 #define PPC_MTOCRF     stringify_in_c(mtcrf)
+#define PPC_LR_STKOFF  4
 
 #endif
 
index 2d29752..b485a87 100644 (file)
@@ -127,3 +127,31 @@ _GLOBAL(__setup_cpu_power7)
 _GLOBAL(__restore_cpu_power7)
        /* place holder */
        blr
+
+#ifdef CONFIG_EVENT_TRACING
+/*
+ * Get a minimal set of registers for our caller's nth caller.
+ * r3 = regs pointer, r5 = n.
+ *
+ * We only get R1 (stack pointer), NIP (next instruction pointer)
+ * and LR (link register).  These are all we can get in the
+ * general case without doing complicated stack unwinding, but
+ * fortunately they are enough to do a stack backtrace, which
+ * is all we need them for.
+ */
+_GLOBAL(perf_arch_fetch_caller_regs)
+       mr      r6,r1
+       cmpwi   r5,0
+       mflr    r4
+       ble     2f
+       mtctr   r5
+1:     PPC_LL  r6,0(r6)
+       bdnz    1b
+       PPC_LL  r4,PPC_LR_STKOFF(r6)
+2:     PPC_LL  r7,0(r6)
+       PPC_LL  r7,PPC_LR_STKOFF(r7)
+       PPC_STL r6,GPR1-STACK_FRAME_OVERHEAD(r3)
+       PPC_STL r4,_NIP-STACK_FRAME_OVERHEAD(r3)
+       PPC_STL r7,_LINK-STACK_FRAME_OVERHEAD(r3)
+       blr
+#endif /* CONFIG_EVENT_TRACING */